Machine Learning and AI methods for Design Of Experiments in the study of electrical contact

Abstract

The increase in contact resistance within battery systems remains poorly understood, despite being one of the primary contributors to battery degradation. This project aims to investigate this phenomenon by modeling the electrical interface as a contact between a pin and a busbar. The main goal is to build a DOE and machine learning framework, using Bayesian optimization to efficiently explore the parameter space.
Experimental data showed a hysteresis in the resistance behavior, but the available inputs were insufficient to explain it. To address this, a set of simulations was performed using Ansys to generate complementary data. In particular, two additional features were extracted: the vertical deformation of the pin and the contact surface area between the pin and the busbar. These were added in an attempt to improve model interpretability and capture the underlying physical mechanisms.
In the final step, symbolic regression using PySR was used to derive an explicit equation characterizing the contact resistance, as a more interpretable alternative to black-box models like neural networks.
The methodology was successfully implemented. However, the results of the symbolic regression were not fully convincing, likely due to a lack of relevant features in the dataset or insufficient physical representation. The models struggled to converge to meaningful expressions that could robustly explain the evolution of the resistance.
Finally, while the project lacks maturity in its current form, it shows promising directions.
Additional work is needed to identify and integrate more relevant physical parameters,such as oxidation effects. Still, the DOE and machine learning pipeline proved effective and general enough to be reused in other applications involving physical modeling,especially where symbolic interpretability is a priority.

Popular Abstract

The degree project aims to develop new methods for Design of Experiments
using Machine Learning algorithms.
You probably wonder why your battery keeps degrading. Few people know this, but a key reason is the increase in electrical resistance within screw joints, the connections between components. Studying these joints is challenging, as many variables can affect resistance.
Experiments are therefore expensive and time-consuming. This thesis aims to simplify the study by using machine learning to design and conduct experiments more efficiently.
To establish electrical contact, pressure is needed to break the oxide layer. The focus is
therefore on this contact force. To simplify the setup, a pin-to-busbar contact is used.
A loading cycle, repeated twice, is applied to the tip, and resistance is measured as current flows through the contact. Three experiments are carried out: one with copper-on-copper, one with a Semi-Bright nickel coating, and one with a Sulfamate nickel coating.
Initially, only Force and Resistance are measured. A hysteresis appears: the second cycle doesn’t give the same resistance as the first, due to plastic deformation. But this can’t be explained with only two variables. So the system is modeled in Ansys to add two more: Contact Area and tip Deformation.
In parallel, the core of the project is to apply machine learning to guide experiments.
Instead of planning all tests in advance, an initial seed is used. The algorithm learns from it and selects new experiments by choosing the points with the highest uncertainty.
The method used is called Gaussian Process, a Bayesian technique that estimates the best-fitting curve and the confidence in its predictions.
Finally, another AI method is tested: Symbolic Regression. Unlike neural networks, it finds a mathematical equation to describe the dataset. Based on Genetic Programming, it works like evolution and provides interpretable results, ideal for physics. With the additional variables from Ansys, it was possible to apply the algorithm to model the contact.
In the end, the dataset wasn’t rich enough to fully validate the model, but promising results appeared in a simplified case. The Design of Experiments method worked well, though not applied fully. This project lays the foundation for future work using AI to guide experimental research.

• author: Moryousef, Nathaniel
• supervisor: Rixin Yu
• organization: Department of Energy Sciences
• course: MVKM01 20251
• year: 2025
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Machine Learning, Symbolic Regression, Design of Experiment, Electrical Contact
• report number: ISRN LUTMDN/TMPH-25/5622-SE
• ISSN: 0282-1990
• language: English
• id: 9199108
• date added to LUP: 2025-06-16 13:34:55
• date last changed: 2025-06-16 13:34:55